---
title: "Where Does Demographic Capital Go? Bilateral Evidence from a Gravity Model"
author: "Working Paper"
date: "February 2026"
abstract: |
  We study whether bilateral demographic differences predict the direction and magnitude of international capital flows. Using bilateral portfolio investment positions from the IMF's Coordinated Portfolio Investment Survey (104,965 country-pair-year observations across 7,885 pairs) and direct investment positions from the Coordinated Direct Investment Survey (109,647 observations across 11,706 pairs), we estimate augmented gravity models that add bilateral demographic distance---the difference in Fair-Dominguez polynomial age-structure variables between origin and destination---to standard gravity regressors (distance, contiguity, common language, colonial ties, GDP). Bilateral demographic distance is highly significant for portfolio investment (all p < 0.001): countries with older populations hold substantially larger portfolio positions in younger countries, with the effect concentrated in debt securities rather than equity. The effect is amplified by destination-country financial openness (all KAOPEN interactions p < 0.023). By contrast, foreign direct investment does not respond to demographic distance (all p > 0.37), consistent with FDI being driven by production networks rather than savings. A two-stage statistical decomposition using fitted bilateral yield differentials shows that 58 percent of the bilateral demographic R² improvement is associated with rate differentials, with 42 percent attributable to non-rate channels---a larger rate-associated share than the 9 percent found in multilateral current account decompositions, reflecting the bilateral model's ability to capture gross yield-seeking behavior. A companion channel decomposition on 140 countries confirms that 88 percent of the net multilateral effect is direct. The gravity results are robust to excluding Central Asia and Caucasus country pairs (coefficient change < 7 percent), and a leave-one-region-out jackknife shows stability across 9 of 11 regions. Projections through 2050 using UN population forecasts, WEO GDP growth, and a general equilibrium rate clearing overlay identify Korea, China, and Southern Europe as the next wave of demographic capital exporters; the Korea--India pair is projected to nearly triple in bilateral portfolio holdings even after GE damping. These bilateral results provide sharper identification than multilateral current account regressions: the demographic distance mechanism survives pair-level analysis and is not driven by the country groups whose sensitivity has been documented in prior work.
keywords: "gravity model, bilateral capital flows, demographics, portfolio investment, lifecycle hypothesis, CPIS"
jel: "F21, F32, J11, G15"
bibliography: references.bib
---

# 1. Introduction

A growing literature documents that demographic structure predicts current account balances: countries with large working-age cohorts tend to run surpluses, while those with dependent-heavy populations run deficits [e.g., @higgins1998; @koomen2020]. However, this evidence relies on multilateral regressions---regressing each country's current account on its own demographic variables. Such specifications face identification challenges: coefficient instability across country samples, sensitivity to particular country groups such as the Central Asia and Caucasus (CCA) transition economies, and the difficulty of distinguishing demographic effects from correlated institutional or structural factors.

This paper takes a different approach. Instead of asking whether a country's own demographics predict its current account, we ask: **does the demographic *difference* between two countries predict the bilateral capital flow between them?** This bilateral specification provides sharper identification for three reasons. First, pair-level variation is far richer than country-level variation: with $N$ countries we have $N(N-1)$ pairs, and the cross-section identifies the demographic effect from within-year variation across thousands of origin-destination combinations. Second, the gravity framework provides a well-established set of controls for bilateral frictions (distance, language, colonial ties), reducing omitted variable concerns. Third, the bilateral specification directly tests the mechanism implied by the lifecycle hypothesis: if aging countries generate excess savings and young countries offer investment opportunities, we should observe directed flows from old to young.

We combine this bilateral evidence with a channel decomposition analysis. Using two-stage mediation regressions on the multilateral panel of 140 countries, we decompose the demographic current account effect into four channels: real exchange rate adjustment, fiscal policy response, interest rate transmission, and a residual direct channel. The decomposition reveals that approximately 88 percent of the demographic effect operates through direct quantity adjustment---savings and investment responding directly to age structure---rather than through price mechanisms. This finding provides a strong prior for the gravity results: if the demographic effect is primarily direct rather than price-associated, bilateral demographic distance should predict flows even after controlling for price variables.

Our results confirm this prediction. Using portfolio investment positions from the IMF's Coordinated Portfolio Investment Survey (CPIS) across 7,885 country pairs from 2001--2024, we find that all three bilateral demographic distance variables are highly significant (all p < 0.001). The effect is economically meaningful: a one-standard-deviation increase in demographic distance raises bilateral portfolio holdings by approximately 15--20 percent. The demographic effect is concentrated in debt securities rather than equity, consistent with aging populations allocating savings to fixed-income instruments. Foreign direct investment, by contrast, shows no response to demographic distance (all p > 0.37), consistent with FDI being driven by production networks and market access rather than savings allocation.

Financial openness amplifies the bilateral demographic flow. All three interactions between demographic distance and destination-country KAOPEN are significant (p < 0.023), indicating that capital follows demographic gradients more readily when the recipient country's capital account is open. This bilateral finding mirrors the multilateral result from the companion 140-country paper, where KAOPEN interactions are significant for developing countries but not for advanced economies already at the openness ceiling.

The gravity results are robust to the exclusion of CCA country pairs---the group identified as the sensitivity tipping point in multilateral analysis. Dropping all CCA-involved pairs changes the demographic distance coefficients by less than 7 percent, and all remain significant at p < 0.001. A leave-one-region-out jackknife shows coefficient stability across 9 of 11 regions, with East Asia and the Middle East and North Africa as the only sensitive exclusions---regions where large demographic gradients (Japan/Korea vs Southeast Asia; Gulf states vs North Africa) drive much of the bilateral variation.

The paper proceeds as follows. Section 2 reviews related literature. Section 3 presents the gravity framework and channel decomposition methodology. Section 4 describes the data. Section 5 presents the main gravity results. Section 6 reports the channel decomposition. Section 7 presents robustness tests. Section 8 projects bilateral flow reallocations through 2050. Section 9 discusses implications. Section 10 concludes.


# 2. Literature Review

## 2.1 Gravity Models of International Finance

The gravity model, originally developed for bilateral trade [@tinbergen1962; @anderson2003], has been extended to international finance by @portes2005, who showed that bilateral equity flows follow gravity patterns: they increase with economic mass and decrease with distance, even after controlling for information and transaction costs. @lane2008 established that bilateral asset holdings also follow gravity, with distance proxying for information asymmetries and monitoring costs. These papers demonstrated that the same frictions governing goods trade also govern financial flows, motivating our use of gravity as the baseline for testing demographic effects.

A key distinction in this literature is between portfolio investment and foreign direct investment. @daude2007 showed that FDI responds more to institutional quality and less to financial market depth, while portfolio flows are more sensitive to information costs and market infrastructure. Our finding that demographics predict portfolio but not FDI flows is consistent with this distinction: portfolio investment is a savings-allocation decision responsive to lifecycle dynamics, while FDI is a production-location decision driven by different factors.

## 2.2 Demographics and Bilateral Capital Flows

The multilateral literature on demographics and current accounts is extensive [see @higgins1998; @koomen2020; @imfeba2013]. However, bilateral tests are scarce. @lane2006 included demographic variables in gravity regressions for bilateral equity holdings but found mixed results, possibly due to limited country coverage and the use of raw demographic ratios rather than the polynomial technique that captures the full age-distribution shape. Our contribution is to bring the Fair-Dominguez polynomial---the state of the art in demographic modeling of external balances---to the gravity framework, with comprehensive bilateral flow data covering both portfolio investment and FDI.

## 2.3 Channel Decomposition

The question of *how* demographics affect external balances has received less attention than *whether* they do. @carvalho2016 propose that demographics compress equilibrium interest rates in aging economies, generating capital outflows toward younger economies with higher returns. @obstfeld2005 emphasize the real exchange rate channel: demographic shifts alter relative prices, affecting competitiveness and thereby current accounts. The fiscal channel---aging populations increasing pension expenditure and reducing fiscal balances---has been documented by @gruber2007.

Our decomposition builds on the mediation analysis framework of @baron1986, adapted for panel data. We estimate the share of the demographic current account effect transmitted through each channel and find that price mechanisms (exchange rates and interest rates combined) account for only about 12 percent, with the remainder operating through direct quantity adjustment of savings and investment. This is consistent with the predictions of overlapping-generations models [@krueger2007; @backus2014], which generate large direct effects and modest price channel transmission.


# 3. Methodology

## 3.1 Gravity Model with Demographic Distance

Our baseline gravity specification is:

$$\ln(\text{Flow}_{ijt}) = \alpha + \beta_1 \ln(d_{ij}) + \beta_2 C_{ij} + \beta_3 L_{ij} + \beta_4 H_{ij} + \beta_5 \ln(Y_i Y_j) + \sum_{k=1}^{3} \gamma_k \Delta Z_{k,ijt} + \delta_t + u_{ijt}$$

where $\text{Flow}_{ijt}$ is the bilateral portfolio (or FDI) position from country $i$ in country $j$ at time $t$; $d_{ij}$ is population-weighted bilateral distance; $C_{ij}$, $L_{ij}$, and $H_{ij}$ are indicators for contiguity, common official language, and colonial ties; $Y_i Y_j$ is the product of nominal GDPs; $\delta_t$ are year fixed effects; and $\Delta Z_{k,ijt} = Z_{k,it} - Z_{k,jt}$ is the bilateral demographic distance in polynomial variable $k$.

The demographic polynomial variables $Z_1$, $Z_2$, $Z_3$ are constructed following @fair1991 and @koomen2020. The population is divided into 17 five-year age groups ($g = 1, \ldots, 17$), and age-group coefficients are constrained to lie on a cubic polynomial: $\alpha_g = \gamma_1 g + \gamma_2 g^2 + \gamma_3 g^3$. The variables $Z_k = \sum_g g^k \cdot s_{g,it}$ aggregate the demeaned population shares $s_{g,it}$ weighted by polynomial terms.

The key hypothesis is $\gamma_k \neq 0$: bilateral demographic distance predicts bilateral flows after controlling for standard gravity variables. If the lifecycle mechanism is operative, we expect $\gamma_1 > 0$: countries that are demographically older than their partners hold larger outward positions.

We estimate this model using pooled GLS with panel-wide AR(1) correction, treating each country pair as the panel entity. This accounts for the high serial correlation in bilateral positions (estimated $\hat{\rho} \approx 0.94$) while allowing the full cross-section to identify the gravity and demographic coefficients.

A natural alternative would be pair fixed effects, which would absorb all time-invariant bilateral characteristics (distance, language, colonial ties) and identify the demographic coefficients solely from within-pair variation over time. We do not adopt this approach for two reasons. First, the demographic polynomial variables change slowly within pairs---the cross-sectional variation across the 7,885 country pairs is far larger than the within-pair variation over 24 years---so pair fixed effects would absorb much of the identifying variation. Second, the time-invariant gravity variables (distance, colonial ties) are substantively important: they capture the bilateral frictions that determine the baseline geography of capital allocation, against which demographic distance operates as a marginal reallocation force. The pooled GLS specification exploits both cross-sectional and time-series variation while the AR(1) correction addresses the resulting serial correlation.

A related concern is multilateral resistance [@anderson2003]: bilateral flows depend not only on bilateral frictions but on the full matrix of alternative opportunities. The structural gravity solution is to include origin$\times$year and destination$\times$year fixed effects, which absorb all country-time-varying characteristics. However, our demographic distance variables $\Delta Z_{k,ijt}$ are constructed from country-level demographic variables, and country$\times$year fixed effects would absorb all variation in the $Z_k$ themselves---leaving only the interaction with time-invariant pair characteristics to identify $\gamma_k$. We instead control for multilateral resistance indirectly through the log GDP product (which proxies for market size and absorbs much of the outward and inward multilateral resistance terms), year fixed effects (which capture global trends), and the rich set of bilateral gravity variables (distance, language, colonial ties) that are the primary source of variation in multilateral resistance across pairs. The PPML robustness check (Section 7.7), which naturally handles the multiplicative structure of the gravity model, confirms that the demographic distance coefficients retain their signs and significance under this alternative estimation approach.

The two approaches identify different estimands. Our preferred pooled specification asks: *does bilateral demographic distance predict bilateral positions, controlling for standard gravity variables?* This is identified from both cross-pair and within-pair variation. A structural gravity specification with origin$\times$year and destination$\times$year FE would instead ask: *does bilateral demographic distance predict bilateral positions conditional on each country's time-varying multilateral characteristics?* Since the $Z_k$ are country-level variables absorbed by country$\times$year FE, the structural gravity estimand reduces to whether demographic distance matters *more for certain pair frictions*---an interaction-based heterogeneity question rather than a test of the baseline demographic distance effect. Section 7.9 presents pair FE results that partially address this by absorbing time-invariant pair characteristics while preserving the country-time variation in demographics.

The AR(1) correction in our PanelGLS estimator addresses serial correlation within pairs ($\hat{\rho} \approx 0.94$), which is the dominant form of dependence in the data. A more conservative approach would cluster standard errors by reporter or partner country to allow for arbitrary within-country correlation across pairs; we present two-way clustered standard errors (by reporter and partner) in Section 7.8 and note that our key findings are confirmed by two independent estimation methods---logistic regression on the extensive margin (Section 7.3) and PPML on levels (Section 7.7)---that use different error structures.

We emphasize that our preferred pooled specification is descriptive and predictive: it documents the strong association between bilateral demographic distance and portfolio positions, controlling for standard gravity variables. We do not claim that demographic distance is exogenous to all confounders; time-invariant country traits correlated with demographics (institutions, financial depth, risk tolerance) could contribute to the estimated coefficients. Pair fixed effects results, presented in Section 7.9, provide a partial guard against such confounding by exploiting only within-pair variation, though at the cost of substantially reduced statistical power.

## 3.2 Extensions

**KAOPEN interactions.** We augment the baseline with $\Delta Z_k \times \text{KAOPEN}_j$, testing whether destination-country financial openness amplifies the demographic flow. The destination rather than origin KAOPEN is used because the binding constraint for lifecycle capital flows is the ability to *enter* the younger economy.

**Flow-type decomposition.** We estimate separately for portfolio equity, portfolio debt, and FDI. The lifecycle hypothesis predicts stronger effects for debt (savings-driven) than equity (risk-driven), and weaker effects for FDI (production-driven).

**Price controls.** We add bilateral interest rate differentials to test whether demographic distance effects survive controlling for the price channel. The companion decomposition shows only 12 percent of the demographic CA effect operates through prices, so we expect limited attenuation.

## 3.3 Channel Decomposition Methodology

The statistical decomposition follows a two-stage approach applied to the multilateral panel of 140 countries. We note that this is a descriptive decomposition of correlations, not a causal mediation analysis---the channel variables are endogenous and the shares should be interpreted as accounting identities rather than structural parameters:

**Stage 1** estimates the effect of demographics on each channel variable separately:

$$\text{Channel}_{it} = \alpha^{(1)} + \sum_k \phi_k Z_{k,it} + \beta^{(1)} X_{it} + u^{(1)}_{it}$$

where Channel $\in$ {log REER, fiscal balance/GDP, gross outflows/GDP, gross inflows/GDP}.

**Stage 2** estimates the effect of each channel on the current account:

$$\text{CA}_{it}/\text{GDP}_{it} = \alpha^{(2)} + \theta \cdot \text{Channel}_{it} + \beta^{(2)} X_{it} + u^{(2)}_{it}$$

The indirect effect through each channel is $\phi_k \times \theta$, and the share statistically associated with each channel is:

$$s_{\text{channel},k} = \frac{\phi_k \times \theta}{\gamma_k^{\text{total}}}$$

where $\gamma_k^{\text{total}}$ is the total (reduced-form) demographic effect from the direct CA regression. The residual share $1 - \sum s_{\text{channel}}$ is the direct/quantity channel.


# 4. Data

## 4.1 Bilateral Capital Flows

**Portfolio investment.** Bilateral portfolio investment positions are from the IMF's Coordinated Portfolio Investment Survey (CPIS), accessed through the PIP database. The CPIS reports the stock of portfolio investment assets held by country $i$ in country $j$, disaggregated into equity and debt securities. Coverage spans 2001--2024 with 86 reporting countries and 216 partner countries. We use outward positions (assets) as the dependent variable, yielding 116,926 positive bilateral-year observations for total portfolio investment.

**Direct investment.** Bilateral FDI positions are from the IMF's Coordinated Direct Investment Survey (CDIS), accessed through the DIP database. Coverage begins in 2009 with 219 reporting countries, yielding 119,794 positive bilateral-year observations.

## 4.2 Gravity Variables

Bilateral gravity variables (population-weighted distance, contiguity, common official language, common ethnolinguistic language, colonial ties, and common colonizer) are from the CEPII GeoDist database [@mayer2011], covering 224 countries and 49,952 directed pairs. After merging bilateral flows with gravity variables and restricting to pairs with non-missing distance and demographics, the estimation sample contains 104,965 portfolio observations across 7,885 pairs and 109,647 FDI observations across 11,706 pairs.

## 4.3 Demographics and Controls

Demographic polynomial variables, macroeconomic controls, and the Chinn-Ito financial openness index are drawn from the 140-country panel constructed in the companion paper. We restrict to years $\leq$ 2024 because the demographic panel includes UN population projections through 2101. Demographics are from the UN World Population Prospects 2024 revision. GDP is from the IMF World Economic Outlook. The KAOPEN index follows @chinnito2006. Interest rate data (government bond yields, lending rates) are from the IMF International Financial Statistics and OECD Main Economic Indicators.

Table 1 summarizes coverage.

**Table 1: Bilateral Data Coverage**

| Flow type | Positive obs | Pairs | Reporters | Partners | Years |
|:----------|:------------|:------|:----------|:---------|:------|
| Portfolio total | 116,926 | 9,013 | 86 | 216 | 2001--2024 |
| Portfolio equity | 85,614 | 6,813 | 85 | 216 | 2001--2024 |
| Portfolio debt | 95,020 | 7,952 | 85 | 215 | 2001--2024 |
| FDI outward | 119,794 | 12,904 | 219 | 218 | 2009--2024 |

*Notes:* Positive bilateral-year observations from CPIS/CDIS before merging with gravity variables. The estimation sample is smaller (104,965 portfolio obs across 7,885 pairs; 109,647 FDI obs across 11,706 pairs) after requiring non-missing gravity and demographic variables.


# 5. Gravity Results

## 5.1 Baseline Gravity

Table 2 reports the main estimation results.

**Table 2: Gravity Model Estimates**

| | (2a) Baseline | (2b) + Demographics | (2c) + KAOPEN int. | (2d) Equity | (2d) Debt | (2d) FDI |
|:--|:--:|:--:|:--:|:--:|:--:|:--:|
| log(distance) | -0.919*** | -0.953*** | -0.919*** | -0.794*** | -0.853*** | -0.872*** |
| | (0.032) | (0.032) | (0.032) | (0.042) | (0.031) | (0.028) |
| Contiguity | -1.171*** | -1.089*** | -0.989*** | -0.827*** | -0.781*** | 0.562*** |
| | (0.158) | (0.157) | (0.152) | (0.198) | (0.151) | (0.124) |
| Common language | 1.593*** | 1.499*** | 1.635*** | 1.770*** | 1.091*** | 1.508*** |
| | (0.080) | (0.080) | (0.078) | (0.106) | (0.079) | (0.064) |
| Colonial ties | 0.087 | -0.043 | -0.128 | 0.209 | -0.189 | 1.485*** |
| | (0.154) | (0.153) | (0.147) | (0.192) | (0.148) | (0.130) |
| log(GDP product) | 0.704*** | 0.718*** | 0.728*** | 0.732*** | 0.603*** | 0.631*** |
| | (0.010) | (0.010) | (0.009) | (0.012) | (0.010) | (0.008) |
| $\Delta Z_1$ | | 3.675*** | 3.961*** | 1.728** | 3.881*** | 0.218 |
| | | (0.604) | (0.724) | (0.785) | (0.631) | (0.527) |
| $\Delta Z_2$ | | -0.489*** | -0.523*** | -0.138 | -0.529*** | 0.045 |
| | | (0.086) | (0.103) | (0.111) | (0.089) | (0.076) |
| $\Delta Z_3$ | | 0.019*** | 0.020*** | 0.003 | 0.021*** | -0.003 |
| | | (0.003) | (0.004) | (0.004) | (0.004) | (0.003) |
| KAOPEN$_j$ | | | 0.217*** | | | |
| | | | (0.014) | | | |
| $\Delta Z_1 \times$ KAOPEN$_j$ | | | 1.164*** | | | |
| | | | (0.378) | | | |
| $\Delta Z_2 \times$ KAOPEN$_j$ | | | -0.140*** | | | |
| | | | (0.054) | | | |
| $\Delta Z_3 \times$ KAOPEN$_j$ | | | 0.005** | | | |
| | | | (0.002) | | | |
| R² | 0.232 | 0.240 | 0.288 | 0.205 | 0.199 | 0.299 |
| N | 104,965 | 104,965 | 95,653 | 76,247 | 85,859 | 109,647 |
| Pairs | 7,885 | 7,885 | 7,415 | 5,890 | 7,027 | 11,706 |
| $\hat{\rho}$ | 0.945 | 0.944 | 0.936 | 0.954 | 0.938 | 0.950 |

*Notes:* Pooled GLS with AR(1) correction. Year dummies included but not reported. Standard errors in parentheses. \*\*\* p<0.01, \*\* p<0.05, \* p<0.10. Dependent variable: log bilateral position (USD).

Column 1 establishes that bilateral portfolio positions follow standard gravity patterns: distance reduces holdings, common language increases them, and the GDP product elasticity is approximately 0.7. The negative contiguity coefficient---counterintuitive at first glance---is consistent with @portes2005 and reflects the strong role of financial centers (UK, US, Luxembourg) that attract portfolio investment from distant countries.

Column 2 adds bilateral demographic distance. All three $\Delta Z$ variables are highly significant (all p < 0.001), and a Wald test for joint significance strongly rejects the null $H_0: \gamma_1 = \gamma_2 = \gamma_3 = 0$ ($\chi^2(3) = 101.5$, p < 0.001). The positive $\Delta Z_1$ coefficient indicates that when the origin country is older than the destination, bilateral portfolio holdings are larger---directly consistent with the lifecycle prediction of old-to-young capital flows. Adding demographics improves R² from 0.232 to 0.240.

Column 3 adds KAOPEN interactions. The level effect of destination openness is large and significant (0.217, p < 0.001): open countries receive more portfolio investment. All three demographic-openness interactions are significant, with the $\Delta Z_1 \times \text{KAOPEN}_j$ interaction at 1.164 (p = 0.002). The joint Wald test for the three KAOPEN interactions is $\chi^2(3) = 21.4$ (p < 0.001), and the joint test for all six demographic variables (levels plus interactions) gives $\chi^2(6) = 102.4$ (p < 0.001). This means the demographic flow effect is substantially amplified when the destination country is financially open---capital follows demographic gradients more readily when barriers are absent. R² rises to 0.288.

## 5.2 Portfolio Debt vs. Equity vs. FDI

Columns 4--6 decompose by flow type. The results reveal a sharp hierarchy:

- **Portfolio debt** (column 5) shows the strongest demographic response. All $\Delta Z$ coefficients are highly significant (all p < 0.001), and the joint Wald test rejects decisively ($\chi^2(3) = 108.6$, p < 0.001). Magnitudes are close to the total portfolio results, consistent with aging populations channeling lifecycle savings into fixed-income instruments.

- **Portfolio equity** (column 4) shows a weaker response. Only $\Delta Z_1$ is significant (p = 0.028); $\Delta Z_2$ and $\Delta Z_3$ are not. The joint test is marginal ($\chi^2(3) = 6.9$, p = 0.076). Equity investment is driven more by risk appetite and growth expectations than by savings allocation.

- **FDI** (column 6) shows no demographic response whatsoever (all p > 0.37). The joint test confirms the null ($\chi^2(3) = 1.3$, p = 0.72). Foreign direct investment is driven by production networks, market access, natural resources, and institutional quality---not by savings behavior. This null result is itself informative: it rules out the possibility that our demographic distance variables are simply proxying for development-level differences or institutional distance, which would also predict FDI.

The debt-not-equity pattern is consistent with the channel decomposition finding that the interest rate channel accounts for 9 percent of the demographic CA effect. Bond markets are the asset class where demographic savings pressure most directly meets yield-seeking capital.

## 5.3 Two-Stage Rate Decomposition

A key limitation of the direct price control test (Model 2e, discussed in Section 7.4) is that observed bilateral bond yield differentials are available for only 506 OECD pairs. To test the rate channel on the full sample, we use a two-stage approach following @carvalho2016. In the first stage, we estimate the mapping from demographics to bond yields on the 23-country OECD subsample where yields are observed: $\hat{r}_i = \hat{\beta}_1 Z_{1,i} + \hat{\beta}_2 Z_{2,i} + \hat{\beta}_3 Z_{3,i}$, yielding coefficients $\hat{\beta}_1 = 16.3$, $\hat{\beta}_2 = -2.07$, $\hat{\beta}_3 = 0.072$ (individually insignificant at p = 0.12--0.18, but jointly capturing 2.7 percent of yield variation). In the second stage, we apply these coefficients to construct fitted bilateral rate differentials $\Delta\hat{r}_{ij} = \hat{r}_i - \hat{r}_j$ for all country pairs, and estimate the gravity model with $\Delta\hat{r}_{ij}$ as the sole demographic variable.

**Table 2b: Rate Mediation Test**

| | Model 2b: Full $\Delta Z$ | Model 2f: Fitted $\Delta\hat{r}$ only |
|:--|:--:|:--:|
| log(distance) | -0.953*** | -0.923*** |
| | (0.032) | (0.032) |
| $\Delta Z_1$ | 3.675*** | |
| | (0.604) | |
| $\Delta Z_2$ | -0.489*** | |
| | (0.086) | |
| $\Delta Z_3$ | 0.019*** | |
| | (0.003) | |
| Fitted $\Delta\hat{r}_{ij}$ | | -0.161*** |
| | | (0.020) |
| R² | 0.240 | 0.237 |
| N | 104,965 | 104,965 |

*Notes:* Both models include full gravity controls (contiguity, common language, colonial ties, log GDP product) and year dummies. Standard errors in parentheses. \*\*\* p<0.01.

The fitted rate differential is highly significant ($\beta = -0.161$, p < 0.001): when the origin country has higher demographically-predicted yields (typically younger), bilateral holdings are lower, consistent with capital flowing from low-yield aging countries to high-yield younger countries.

The R²-based decomposition compares incremental fit over the baseline gravity model:

- Full $\Delta Z$ model: $\Delta R^2 = +0.0086$ (all demographic channels)
- Fitted $\Delta\hat{r}$ only: $\Delta R^2 = +0.0050$ (rate-associated component)
- **Rate-associated share: 58%** of the bilateral demographic R² improvement
- **Non-rate (direct) share: 42%**

This exercise is best read as answering: *how much of the incremental fit from demographics is also captured by demographically-predicted yield differentials?* It is not a valid IV strategy and should not be interpreted as a causal mediation decomposition. The fitted-rate regression is a diagnostic consistent with a yield-seeking mechanism: 58 percent of what demographics explain about bilateral portfolio positions is also explained by the yield differential that demographics predict.

Importantly, the residual 42 percent confirms that a substantial portion of the demographic bilateral association operates through non-rate channels---direct savings allocation driven by age structure---even in the bilateral specification. Demographics predict bilateral portfolio positions above and beyond their association with yield differentials.


# 6. Channel Decomposition

## 6.1 Clearing Budget

Table 3 reports the statistical decomposition from the two-stage analysis on the 140-country multilateral panel. We emphasize that these shares reflect correlational patterns consistent with each channel, not causally identified transmission pathways.

**Table 3: Demographic CA Effect — Channel Decomposition**

| Channel | Full sample | Advanced economies | Emerging/developing |
|:--------|:-----------|:-------------------|:-------------------|
| Real exchange rate | 3.5% | 23.4% | 0.2% |
| Fiscal policy | -1.8% | -0.1% | -1.0% |
| Interest rates | 9.0% | 9.0% | 9.0% |
| **Direct/residual** | **89.3%** | **67.6%** | **91.8%** |

*Notes:* Shares computed from two-stage mediation analysis. Interest rate share from GE clearing model. Negative fiscal share indicates fiscal policy weakly offsets demographic CA pressure.

The dominant finding is that nearly 90 percent of the demographic effect on current accounts operates through a direct channel that does not pass through observable price variables. This is not a puzzle---it is the lifecycle mechanism in action. Overlapping-generations models predict that demographic structure directly affects aggregate saving rates through age composition (young dependents consume, prime-age adults save, retirees dissave), shifting the savings-investment balance without requiring exchange rate or fiscal intermediation. Calibrated OLG models by @krueger2007 and @backus2014 predict price channel shares of 10--20 percent, consistent with our empirical estimate of approximately 12 percent.

## 6.2 Advanced vs. Emerging Economies

The decomposition reveals a striking difference between advanced and emerging economies. In advanced economies, the real exchange rate channel absorbs 23 percent of the demographic effect---a meaningful share reflecting well-functioning foreign exchange markets and flexible exchange rate regimes. In emerging and developing economies, the REER channel collapses to near zero (0.2%), likely due to managed exchange rates, capital controls, and underdeveloped FX markets that prevent real exchange rates from reflecting demographic fundamentals.

Financial openness gates the REER channel: interactions between demographic variables and KAOPEN significantly predict REER movements ($Z_2 \times \text{KAOPEN}$: p = 0.027; $Z_3 \times \text{KAOPEN}$: p = 0.018), but the same interactions do not predict gross capital flows (all p > 0.19). The REER channel operates in open economies; in closed economies, neither REER nor flow channels transmit the demographic effect.

## 6.3 Connecting Bilateral and Multilateral Channels

The bilateral gravity results and the multilateral channel decomposition measure different objects, but they tell a coherent story. The multilateral decomposition asks: of the total demographic effect on a country's *net* current account, how much passes through observable price variables? The answer is approximately 12 percent (3.5% REER + 9% interest rates), with 88 percent operating through direct savings-investment adjustment. The bilateral two-stage analysis (Section 5.3) asks a different question: of the bilateral demographic effect on *gross* portfolio positions between country pairs, how much is explained by yield-seeking? The answer is 58 percent.

The difference---58 percent bilateral versus 9 percent multilateral---is not contradictory; it follows from the nature of the dependent variables. The multilateral current account is a *net* object ($S - I$): yield-seeking portfolio reallocations across dozens of partners largely wash out in the aggregate. Japan allocates bonds to the US, Australia, and Brazil on yield grounds, but these gross flows substantially net out, leaving a current account that primarily reflects *how much* Japan saves---a direct lifecycle effect. The bilateral CPIS position is a *stock allocation* object that can respond strongly to relative yields even when the net CA barely moves. Thus the "rate-associated share" is not expected to match across the two: they are different dependent variables measuring different equilibrium objects. In short, interest rates are a major determinant of the *geography* of demographic capital flows (bilateral) but a modest determinant of their *aggregate magnitude* (multilateral).

The debt-not-equity finding reinforces this interpretation. Demographics predict bilateral debt holdings (all $\Delta Z$ coefficients p < 0.001) but not equity (only $\Delta Z_1$ significant) or FDI (all null). The interest rate channel operates through bond markets---demographics compress yields in aging countries, directing bond purchases toward higher-yielding younger economies. This is exactly the asset class where the bilateral rate share should be largest, and the FDI null confirms that the demographic distance variables are not merely proxying for development-level differences.

The practical implication is that the multilateral "88 percent direct" finding should not be interpreted as meaning interest rates are irrelevant. They are highly relevant for *where* demographic capital goes, even though they explain little of *how much* leaves. Policy interventions that alter yield differentials (e.g., quantitative easing, capital controls affecting bond returns) can meaningfully redirect bilateral flows without substantially changing aggregate current account balances.


# 7. Robustness

## 7.1 CCA Sensitivity

The companion 140-country paper documented that the CCA group of transition economies acts as a tipping point for statistical significance in multilateral regressions. Table 4 tests whether this sensitivity extends to bilateral analysis.

**Table 4: CCA Robustness**

| Specification | $\Delta Z_1$ | $\Delta Z_2$ | $\Delta Z_3$ | N |
|:---|:--:|:--:|:--:|:--:|
| Full sample | 3.675*** | -0.489*** | 0.019*** | 104,965 |
| Excluding CCA pairs | 3.877*** (+5.5%) | -0.520*** (-6.3%) | 0.020*** (+6.8%) | 99,459 |
| Excluding CCA non-commodity | 3.683*** (+0.2%) | -0.490*** (-0.2%) | 0.019*** (+0.5%) | 102,819 |

*Notes:* Percentage changes relative to full sample in parentheses.

Excluding all CCA-involved pairs changes coefficients by less than 7 percent, and all remain highly significant (p < 0.001). This is a markedly different result from the multilateral analysis, where CCA exclusion substantially weakened the demographic coefficients. The bilateral gravity framework is inherently more robust because the identification comes from the full cross-section of country pairs, not just from the within-country time series of a few unusual economies.

## 7.2 Leave-One-Region-Out Jackknife

Table 5 summarizes the jackknife analysis.

**Table 5: Jackknife Coefficient Stability ($\Delta Z_1$)**

| Excluded region | Coefficient | p-value | Significant? |
|:---|:--:|:--:|:--:|
| Full sample | 3.675 | <0.001 | Yes |
| Advanced Europe | 4.972 | <0.001 | Yes |
| EU New Members | 3.263 | <0.001 | Yes |
| East Asia | 0.163 | 0.819 | No |
| Southeast Asia | 3.467 | <0.001 | Yes |
| South Asia | 3.689 | <0.001 | Yes |
| Latin America | 4.264 | <0.001 | Yes |
| Middle East & N. Africa | -0.656 | 0.422 | No |
| Sub-Saharan Africa | 5.750 | <0.001 | Yes |
| CCA | 3.877 | <0.001 | Yes |
| Other Europe & CIS | 3.737 | <0.001 | Yes |
| Anglo-Saxon & Pacific | 4.100 | <0.001 | Yes |

*Notes:* Jackknife results for $\Delta Z_2$ and $\Delta Z_3$ show qualitatively similar patterns; full results in supplementary materials.

The jackknife reveals that 9 of 11 regions leave the demographic distance effect significant. The two sensitive exclusions are East Asia and Middle East & North Africa. Both regions contain large demographic gradients---Japan and Korea (very old) investing in Southeast Asia and the Gulf states; Gulf states (young with petrodollar surpluses) investing globally. Their removal eliminates a substantial portion of the identifying variation, which is expected: the demographic distance hypothesis predicts that these regions should contribute disproportionately to the bilateral effect. Unlike the CCA sensitivity in multilateral analysis (which raised concerns about a small group of outliers driving results), the East Asia and MENA sensitivity reflects the *substantive content* of the hypothesis.

## 7.3 Extensive vs. Intensive Margin

We decompose the demographic effect into the extensive margin (does a bilateral position exist?) and the intensive margin (how large is it?).

**Extensive margin.** A logit model on the indicator for positive portfolio holdings shows that all $\Delta Z$ variables significantly predict whether a bilateral position exists (all p < 0.001, pseudo-R² = 0.262). Demographics predict not just how much capital flows, but whether any connection exists at all.

**Intensive margin.** Conditional on a positive position, the GLS estimates on log(flow) replicate the full-sample results (Table 2, column 2). The demographic effect operates on both margins.

## 7.4 Expanded Bond Yield Coverage

A potential concern with Model 2e is the limited sample (506 OECD pairs). We expand the bond yield panel from 23 to 35 countries by sourcing additional OECD MEI long-term interest rate series for Poland, Czech Republic, Hungary, Israel, Chile, South Africa, Iceland, Luxembourg, Slovakia, Slovenia, Russia, and China. This more than doubles the price-control sample from 11,473 to 24,431 observations (1,180 pairs).

**Table 5b: Expanded Price Control Test**

| | 2e original (23 ctry) | 2e expanded (35 ctry) |
|:--|:--:|:--:|
| $\Delta Z_1$ | -6.727*** (0.005) | -3.901** (0.040) |
| $\Delta Z_2$ | 1.124*** (0.000) | 0.589** (0.018) |
| $\Delta Z_3$ | -0.046*** (0.000) | -0.022** (0.015) |
| Rate differential | -0.004 (0.249) | 0.004 (0.102) |
| R² | 0.579 | 0.413 |
| N | 11,473 | 24,431 |
| Pairs | 506 | 1,180 |

*Notes:* p-values in parentheses after coefficients.

The bilateral rate differential remains insignificant even with doubled coverage (p = 0.102), and its sign reverses from negative to positive. The $\Delta Z$ coefficients retain significance but with reversed signs relative to the full sample---a pattern already present in the original 23-country subsample. This sign reversal reflects that the yield-available sample consists predominantly of advanced economies where the demographic-flow relationship differs from the global cross-section: within AE pairs, older countries hold *less* in each other (home bias) rather than more, while the full-sample positive coefficient reflects old-to-young cross-development flows.

We also re-estimate the first-stage Carvalho regression (demographics $\to$ bond yields) on the expanded 35-country panel. The S1 coefficients become $\hat{\beta}_1 = 15.1$ (p = 0.20), $\hat{\beta}_2 = -2.01$ (p = 0.21), $\hat{\beta}_3 = 0.074$ (p = 0.21)---marginally improved p-values relative to the 23-country estimates but still far from conventional significance, with R² declining from 0.019 to 0.006. The original 23-country OECD S1 coefficients produce the correct (negative) sign for fitted rate differentials in the gravity model; the expanded 35-country coefficients reverse the sign (Model 2f-exp: coefficient = +0.235, p < 0.001), consistent with the demographics-yield relationship being primarily an OECD phenomenon. The mediation decomposition using expanded S1 coefficients produces a negative rate-channel share (-13%), confirming that the Carvalho rate channel mechanism should not be extrapolated beyond advanced economies with deep bond markets.

The expansion thus reinforces rather than weakens our core finding: observed bilateral yield differentials do not absorb the demographic effect, even with substantially more data. The two-stage Carvalho approach using OECD-estimated coefficients (Section 5.3) remains the appropriate rate-association diagnostic.

An important implication is that our 58 percent bilateral rate-mediation share (Section 5.3) should be understood as conditional on OECD-level bond market maturity. The first-stage demographics-to-yield mapping is estimated on 23 advanced economies with deep, liquid sovereign bond markets where demographic savings pressure plausibly transmits to equilibrium yields. In economies with shallow or repressed bond markets, the rate channel may be substantially weaker or absent---consistent with the expanded-sample sign reversal documented above. The bilateral demographic effect itself is global (all $\Delta Z$ coefficients remain p < 0.001 on the full 105,000-observation sample), but the share of that effect operating through interest rates versus direct savings allocation likely varies with financial market depth. Readers should interpret the 58/42 rate-versus-direct decomposition as characterizing the transmission mechanism in advanced economies, not as a universal ratio.

## 7.5 Financial Center Robustness

A potential concern is that portfolio transit hubs---jurisdictions where investment is booked for tax or regulatory reasons rather than reflecting genuine bilateral savings allocation---could distort the demographic distance coefficients. Table 5b-ii tests this by excluding financial centers from both reporter and partner sides.

**Table 5b-ii: Financial Center Exclusion**

| | Full sample | Narrow exclusion | Broad exclusion |
|:--|:--:|:--:|:--:|
| $\Delta Z_1$ | 3.675*** | 3.455*** (-6%) | 4.395*** (+20%) |
| $\Delta Z_2$ | -0.489*** | -0.505*** (-3%) | -0.645*** (-32%) |
| $\Delta Z_3$ | 0.019*** | 0.021*** (+11%) | 0.027*** (+41%) |
| R² | 0.240 | 0.352 | 0.320 |
| N | 104,965 | 84,717 | 65,112 |
| Pairs | 7,885 | 6,464 | 5,268 |

*Notes:* Narrow exclusion drops 11 offshore/pass-through jurisdictions (LUX, IRL, CYM, BMU, BHS, PAN, VGB, BHR, MUS, MLT, CYP). Broad exclusion additionally drops 6 major financial hubs (HKG, SGP, CHE, NLD, BEL, GBR). Percentage changes vs. full sample in parentheses.

All demographic distance coefficients remain highly significant (p < 0.001) under both exclusions. Notably, the coefficients are *larger* after excluding financial centers, particularly under the broad exclusion---consistent with transit hubs introducing noise rather than signal. The R² improvement from 0.240 to 0.352 (narrow) suggests that removing offshore pass-through positions reduces measurement error in the dependent variable. The KAOPEN interaction $\Delta Z_1 \times \text{KAOPEN}_j$ weakens under the broad exclusion (from p = 0.002 to p = 0.098), which is expected given that the broad set removes major open-economy financial centers (Singapore, Hong Kong, Switzerland, UK) that exemplify the openness amplification mechanism.

## 7.6 Bootstrapped Clearing Shares

The multilateral channel decomposition is validated with 500 block-bootstrap iterations. The median residual (direct) share is 88.6 percent for $Z_1$, with a 95 percent confidence interval of [25%, 137%]. While the REER and fiscal shares are individually noisy (confidence intervals spanning zero), the residual share is robustly bounded away from zero: even in the worst-case bootstrap draw, at least 25 percent of the demographic effect is direct. The bilateral R²-based mediation decomposition (58% rate / 42% direct from Section 5.3) is a point estimate without analogous uncertainty quantification; readers should interpret it as indicative of a substantially larger rate channel at the bilateral level rather than as a precise share.

## 7.7 PPML Robustness

Our baseline estimates use OLS on log bilateral positions, which drops zero-valued observations and may introduce bias from log-linearization under heteroskedasticity [@santos2006]. Table 5c reports Poisson Pseudo-Maximum Likelihood (PPML) estimates on levels, which naturally include zeros and are consistent under heteroskedasticity.

**Table 5c: PPML vs. OLS Comparison (Demographic Coefficients)**

| | OLS 2b | PPML 2b | OLS 2c | PPML 2c |
|:--|:--:|:--:|:--:|:--:|
| $\Delta Z_1$ | 3.675*** | 1.370*** | 3.961*** | -2.519*** |
| $\Delta Z_2$ | -0.489*** | -0.201*** | -0.523*** | 0.404*** |
| $\Delta Z_3$ | 0.019*** | 0.008*** | 0.020*** | -0.017*** |
| KAOPEN$_j$ | | | 0.217*** | 0.433*** |
| $\Delta Z_1 \times$ KAOPEN$_j$ | | | 1.164*** | 3.693*** |
| $\Delta Z_2 \times$ KAOPEN$_j$ | | | -0.140*** | -0.498*** |
| $\Delta Z_3 \times$ KAOPEN$_j$ | | | 0.005** | 0.019*** |
| N (incl. zeros) | 104,965 | 116,184 | 95,653 | 103,736 |
| Zeros | 0 | 63,781 | 0 | 55,889 |

*Notes:* PPML estimated on 50% random subsample for computational feasibility. OLS dependent variable is log(bilateral position); PPML dependent variable is level. Standard Poisson SEs are reported but likely understated due to overdispersion; significance should be interpreted qualitatively.

The PPML results confirm the OLS findings in two important respects. First, in Model 2b (without openness interactions), all three demographic distance variables retain their signs and significance, with PPML magnitudes approximately 37 percent of the OLS estimates---attenuation that is typical when moving from log-linear to level specifications. Second, in Model 2c, the KAOPEN interactions are substantially *larger* under PPML, with the base $\Delta Z$ terms reversing sign. This pattern implies that under PPML, the demographic flow effect operates almost entirely through financially open destinations: at the KAOPEN ceiling (2.28), the total PPML effect of $\Delta Z_1$ is $-2.52 + 3.69 \times 2.28 = +5.9$, strongly positive and larger than the OLS estimate. At closed economies (KAOPEN $\approx$ 0), the effect is negligible or reversed. This is economically sensible: bilateral portfolio positions in closed economies are near zero regardless of demographics.

The PPML standard errors are unrealistically small (z-statistics in the hundreds), reflecting the severe overdispersion inherent in bilateral portfolio data that spans several orders of magnitude. Robust (sandwich) standard errors or quasi-Poisson specifications would be more appropriate for formal inference; the present PPML estimates serve as a specification check confirming that the OLS results are not driven by the log transformation or zero-dropping.

## 7.8 Clustered Standard Errors

Our baseline GLS estimates correct for within-pair serial correlation via AR(1) transformation but assume independence across pairs. In gravity data, pairs sharing a common reporter or partner may exhibit correlated errors---for example, a crisis in reporter $i$ affects all of $i$'s outward positions simultaneously. We present three clustering approaches using OLS (without AR(1) correction) to assess inference sensitivity: pair-clustered (robust to arbitrary within-pair serial correlation; 7,415--7,885 clusters), reporter-clustered (robust to within-reporter correlation across partners; 82 clusters), and two-way clustered by reporter and partner following @cameron2011 ($V_{\text{two-way}} = V_{\text{reporter}} + V_{\text{partner}} - V_{\text{pair}}$).

**Table 5d: Clustered Standard Errors (Model 2c)**

| Variable | GLS SE | Pair SE | Reporter SE | Two-way SE |
|:--|:--:|:--:|:--:|:--:|
| $\Delta Z_1$ | 0.724 | 1.245 | 3.521 | 3.819 |
| $\Delta Z_2$ | 0.103 | 0.174 | 0.498 | 0.533 |
| $\Delta Z_3$ | 0.004 | 0.007 | 0.019 | 0.021 |
| $\Delta Z_1 \times$ KAOPEN$_j$ | 0.378 | 0.693 (***) | 0.891 (***) | 1.249 (**) |
| $\Delta Z_2 \times$ KAOPEN$_j$ | 0.054 | 0.098 (***) | 0.127 (***) | 0.173 (**) |
| $\Delta Z_3 \times$ KAOPEN$_j$ | 0.002 | 0.004 (***) | 0.005 (***) | 0.007 (**) |

*Notes:* OLS estimates (no AR(1) correction). Pair-clustered: 7,415 clusters; reporter-clustered: 82 clusters; two-way: Cameron-Gelbach-Miller. Significance stars for KAOPEN interactions only.

The table reveals a clear hierarchy. Pair-clustered SEs (which handle within-pair serial correlation non-parametrically) are 1.7--1.9$\times$ the GLS SEs---a modest inflation confirming that the GLS AR(1) correction captures most of the within-pair dependence. Reporter-clustered SEs are 2.3--4.9$\times$ the GLS SEs, reflecting the additional correlation across pairs sharing a reporter. Two-way clustering is the most conservative at 3.3--5.3$\times$.

Under all three clustering approaches, the level $\Delta Z$ terms become insignificant (pair-clustered p $\approx$ 0.28--0.45; reporter- and two-way p > 0.67). This reflects the slowly evolving nature of demographic variables: OLS without AR(1) correction leaves massive serial correlation in the errors, inflating clustered SEs well beyond what a properly specified serial correlation model produces.

The most important result is that the **KAOPEN interaction terms---the policy-relevant finding---remain highly significant under all clustering approaches**: pair-clustered (all p < 0.001), reporter-clustered (all p $\leq$ 0.003), and two-way (all p < 0.04). The interaction terms are more robust because they vary with destination-country openness, providing cross-sectional variation beyond the within-reporter and within-partner dimensions that clustering absorbs.

To state this explicitly: the most conservative two-way clustering eliminates significance of the level $\Delta Z$ terms, but our strongest and most policy-relevant inference---that financial openness of the destination amplifies the bilateral demographic flow---survives even this conservative benchmark. This inference is further corroborated by two independent estimation methods with different error structures: logistic regression on the extensive margin (Section 7.3) and PPML on levels (Section 7.7).

## 7.9 Pair Fixed Effects

Our pooled specification identifies demographic distance from both cross-pair and within-pair variation. To isolate purely within-pair identification---testing whether *changes* in bilateral demographic distance predict *changes* in bilateral positions for the *same pair* over time---we estimate pair fixed effects models with year dummies, using OLS on demeaned (within-pair) data.

**Table 5e: Pair Fixed Effects Estimates**

| | Model 2b (no interactions) | Model 2c (with KAOPEN) |
|:--|:--:|:--:|
| $\Delta Z_1$ | $-1.076$ (0.001) | $+0.987$ (0.017) |
| $\Delta Z_2$ | $+0.025$ (0.566) | $-0.307$ (<0.001) |
| $\Delta Z_3$ | $+0.002$ (0.215) | $+0.016$ (<0.001) |
| KAOPEN$_j$ | --- | $+0.042$ (0.001) |
| $\Delta Z_1 \times$ KAOPEN$_j$ | --- | $-1.554$ (<0.001) |
| $\Delta Z_2 \times$ KAOPEN$_j$ | --- | $+0.236$ (<0.001) |
| $\Delta Z_3 \times$ KAOPEN$_j$ | --- | $-0.010$ (<0.001) |
| R² (within) | 0.204 | 0.203 |
| N | 104,965 | 95,653 |
| Pairs | 7,885 | 7,415 |

*Notes:* OLS with pair fixed effects and year dummies. P-values in parentheses.

Pair fixed effects produce two notable findings. First, in Model 2b (without interactions), $\Delta Z_1$ reverses sign to negative ($-1.08$, p = 0.001), while $\Delta Z_2$ and $\Delta Z_3$ become insignificant. The sign reversal reflects a home bias dynamic: within a fixed pair, as the reporter ages relative to its partner, it tends to *reduce* bilateral portfolio holdings---consistent with aging-related portfolio retrenchment or declining international financial engagement. This within-pair pattern is interesting but distinct from the cross-pair question of interest (do older-than-partner countries hold more abroad?).

Second, and more importantly, Model 2c with KAOPEN interactions tells a sharply different story. With interactions included, $\Delta Z_1$ is positive and significant ($+0.99$, p = 0.017), and all three KAOPEN interactions are highly significant (all p < 0.001). This means that within pairs, aging of the reporter relative to the partner predicts *increased* bilateral holdings specifically when the destination is financially open. The within-pair R² is 0.20, comparable to the pooled specification.

The pair FE results reinforce the central finding: the bilateral demographic flow mechanism operates through financial openness. The level effect (Model 2b) is cross-sectional in nature and absorbed by pair FE, but the interaction with destination openness---the policy-relevant margin---survives the most demanding identification strategy available in bilateral data.


# 8. Bilateral Flow Projections

Using Model 2c coefficients and UN population projections through 2050, we project how bilateral portfolio flow allocations will shift as countries age at different rates. These projections incorporate three components: (1) evolving bilateral demographic distance, (2) IMF WEO GDP projections through 2030 (with 2030 GDP held constant for 2040--2050), and (3) a general equilibrium clearing overlay that adjusts for the endogenous world interest rate response to global aging. KAOPEN (financial openness) is held at 2024 values.

## 8.1 Projection Framework

The projected change in log bilateral portfolio position from country $i$ to country $j$ has three components:

$$\Delta \ln(\text{Flow}_{ijt}) = \underbrace{\sum_k \hat{\gamma}_k \Delta(\Delta Z_{k,ijt})}_{\text{Demographics}} + \underbrace{\hat{\beta}_5 \Delta \ln(Y_{it} Y_{jt})}_{\text{GDP growth}} + \underbrace{\hat{\delta}_r (-\Delta r^*)}_{\text{GE rate adjustment}}$$

The demographic component uses Model 2c coefficients (including KAOPEN interactions). GDP growth uses WEO projections (India +73%, Nigeria +58%, China +38% by 2030; held at 2030 for subsequent years). The GE adjustment follows the companion paper's clearing model: we compute the GDP-weighted global partial-equilibrium imbalance, solve for the world rate adjustment $\Delta r^*$ that clears the aggregate capital market, and apply it uniformly to all bilateral flows using the bilateral rate semi-elasticity ($\hat{\delta}_r = -0.161$ from Model 2f).

The clearing rate at 2040--2050 is $\Delta r^* \approx -1.7$ percentage points---world bond yields *fall* as global aging compresses returns, dampening all bilateral portfolio positions by approximately 24 percent ($= \exp(0.161 \times 1.7) - 1$). Crucially, because the GE adjustment is uniform across pairs, it affects the *level* of bilateral flows but not their *geography*: the net reallocation pressure between countries is determined entirely by demographics.

## 8.2 Largest Projected Bilateral Shifts

Table 6 reports the top projected increases by 2050, decomposed into demographic, GDP, and GE components.

**Table 6: Top Projected Bilateral Shifts (2050 vs. 2024)**

| Reporter | Partner | Demo %$\Delta$ | GDP %$\Delta$ | PE total | GE total | GE damping |
|:---------|:--------|:--:|:--:|:--:|:--:|:--:|
| KOR | ARE | +290% | +39% | +443% | +310% | -133pp |
| CHN | ARE | +201% | +59% | +378% | +261% | -117pp |
| ESP | ARE | +197% | +50% | +347% | +238% | -109pp |
| ITA | ARE | +189% | +41% | +309% | +209% | -100pp |
| KOR | ETH | +156% | +58% | +304% | +206% | -99pp |
| KOR | PHL | +151% | +59% | +298% | +201% | -97pp |
| KOR | NGA | +146% | +55% | +281% | +188% | -93pp |
| KOR | IND | +127% | +65% | +275% | +183% | -92pp |
| CHN | ETH | +100% | +80% | +259% | +172% | -88pp |
| CHN | PHL | +95% | +81% | +253% | +167% | -86pp |

*Notes:* Projections use Model 2c coefficients applied to UN WPP 2024 medium-variant demographic projections and IMF WEO GDP projections through 2030 (held constant thereafter). GE clearing uses $\Delta r^* = -1.74$ pp at 2050. Percentage changes computed as $(\exp(\Delta) - 1) \times 100$.

GDP growth amplifies all projections substantially. India's 73 percent WEO GDP growth by 2030 means that every pair involving India as a destination receives a large GDP boost on top of demographic effects. The GE overlay dampens all pairs by approximately 90--130 percentage points, but even after GE adjustment the top pairs show projected increases of 170--310 percent---a tripling for Korea--UAE and a near-tripling for Korea--India.

## 8.3 Time Paths: Partial vs. General Equilibrium

Table 7 compares PE and GE projected changes for selected pairs.

**Table 7: PE vs. GE Projections (% change vs. 2024)**

| Pair | | 2030 | 2040 | 2050 |
|:-----|:--|:--:|:--:|:--:|
| KOR $\to$ IND | PE | +95% | +180% | +275% |
| | GE | +95% | +113% | +183% |
| CHN $\to$ NGA | PE | +106% | +164% | +238% |
| | GE | +106% | +101% | +156% |
| DEU $\to$ IND | PE | +76% | +103% | +103% |
| | GE | +76% | +54% | +54% |
| JPN $\to$ IND | PE | +92% | +96% | +101% |
| | GE | +92% | +49% | +52% |
| USA $\to$ IND | PE | +92% | +103% | +86% |
| | GE | +92% | +54% | +41% |
| USA $\to$ MEX | PE | +42% | +46% | +30% |
| | GE | +42% | +11% | -2% |

*Notes:* GE projections incorporate global clearing rate adjustment ($\Delta r^* = 0$ at 2030, $-1.7$ pp at 2040--2050). 2030 projections are identical for PE and GE because the clearing rate is negligible in the near term.

Several patterns emerge. At 2030, PE and GE projections are identical because the global capital market is approximately in balance---aggregate aging has not yet reached the point where the rate channel must adjust. By 2040--2050, GE damping is substantial: the Korea--India pair grows by +275% (PE) versus +183% (GE), and the USA--Mexico pair turns slightly negative under GE. Japan and Germany pairs are most affected because they are already aged: their demographic component grows slowly while GE damping accumulates, reducing their projected increase by roughly half.

The 2030 projections deserve special attention as the most reliable, since they use actual WEO GDP data (not frozen estimates). At this horizon, GDP growth is the dominant force: India's rapid GDP growth (+73%) amplifies all India-destination pairs substantially. The 2040--2050 projections should be interpreted with more caution, as they extrapolate 2030 GDP and apply a stylized GE clearing model.

## 8.4 Net Reallocation Pressure and Decomposition

Table 8 decomposes the average projected change in each country's outward bilateral flows into demographic, GDP, and GE components.

**Table 8: Decomposition of 2050 Projected Changes (Average Across Partners)**

| Country | Demographics | GDP growth | GE rate adj. | Total (GE) | Net reallocation |
|:--------|:--:|:--:|:--:|:--:|:--:|
| KOR | +0.73 | +0.31 | -0.28 | +0.76 | +1.46 |
| CHN | +0.47 | +0.43 | -0.28 | +0.62 | +0.95 |
| ESP | +0.43 | +0.38 | -0.28 | +0.53 | +0.87 |
| ITA | +0.38 | +0.32 | -0.28 | +0.42 | +0.80 |
| THA | +0.36 | +0.35 | -0.28 | +0.43 | +0.72 |
| DEU | +0.13 | +0.34 | -0.28 | +0.18 | +0.29 |
| JPN | +0.05 | +0.36 | -0.28 | +0.13 | +0.11 |
| USA | -0.05 | +0.38 | -0.28 | +0.05 | -0.09 |
| IND | -0.15 | +0.59 | -0.28 | +0.16 | -0.26 |
| NGA | -0.09 | +0.53 | -0.28 | +0.16 | -0.28 |
| PHL | -0.22 | +0.55 | -0.28 | +0.05 | -0.44 |
| ARE | -0.58 | +0.43 | -0.28 | -0.44 | -1.26 |

*Notes:* Demographics: change in demographic component of log(Flow) from Model 2c. GDP growth: change from WEO GDP projections through 2030. GE rate adj.: bilateral flow adjustment from global clearing rate ($\Delta r^* = -1.74$ pp × $\hat{\delta}_r = -0.161$). Net reallocation: outward minus received demographic pressure (invariant to GE).

The decomposition reveals that the three forces operate differently. GDP growth is the largest component for all countries (+0.3 to +0.6 log points) but varies little across countries because global GDP is growing roughly in line. The GE adjustment is uniform (-0.28) by construction. **Demographics are the sole driver of cross-country differences in reallocation pressure**, which is why the net reallocation column---ranking which countries gain or lose bilaterally---is identical to the demographic-only projection: Korea (+1.46), China (+0.95), and Spain (+0.87) as the top future exporters; UAE (-1.26), Saudi Arabia (-0.53), and Pakistan (-0.51) as top receivers.

Three implications follow. First, the bilateral reallocation of demographic capital is robust to GE feedback: interest rate adjustment dampens the *volume* of cross-border flows but not their *direction*. Second, GDP growth substantially amplifies the raw demographic projection, especially for pairs involving fast-growing emerging markets. Third, the 2030 projections (where WEO GDP data is actual rather than frozen) provide the most conservative and reliable baseline: at this horizon, Korea--India bilateral portfolio holdings are projected to nearly double even without GE adjustment.


# 9. Discussion

## 9.1 Policy Implications

The finding that demographic capital flows are predominantly direct (88%) rather than price-associated has important policy implications. Exchange rate intervention and interest rate policy can absorb only a modest fraction of the demographic savings imbalance. For aging economies like Japan, Germany, and Korea, the lifecycle saving surplus will persist as long as the age distribution remains tilted toward older cohorts, regardless of exchange rate or monetary policy. For young economies seeking to attract this capital, the binding constraint is financial openness: our gravity results show that KAOPEN significantly amplifies the demographic flow, suggesting that capital account liberalization would channel more demographic savings toward countries that need investment.

## 9.2 Demographics and the Lucas Paradox

The extensive margin result (Section 7.3) carries implications for the "Lucas Paradox"---the observation that capital does not flow from rich to poor countries in the quantities neoclassical theory predicts [@lucas1990]. Our logit estimates show that bilateral demographic distance significantly predicts whether a portfolio investment link exists at all (all $\Delta Z$ coefficients p < 0.001, pseudo-R² = 0.262). Demographics thus help explain not just the reallocation of existing capital flows but the *formation* of new international financial linkages. As populations age unevenly, the extensive margin implies that country pairs with large demographic gradients will forge bilateral connections that did not previously exist---a structural expansion of the global financial network driven by lifecycle savings rather than by convergence in returns to capital. This demographic channel is distinct from the institutional and human capital explanations typically offered for the Lucas Paradox, and suggests that global aging may gradually erode the paradox as demographic savings pressure pushes capital toward younger economies regardless of their institutional quality, provided their capital accounts are open.

## 9.3 The FDI Null

The absence of a demographic effect on FDI is substantively important. It implies that the "demographic dividend" often discussed in development policy---the growth boost from a large young working-age population---does not automatically attract foreign productive investment. FDI responds to production fundamentals (market size, infrastructure, institutions), not to the savings behavior of the source country's aging population. Policy efforts to attract FDI from aging advanced economies should focus on the business environment rather than on demographic complementarity.

## 9.4 Limitations

We distinguish between findings that are robust across specifications and those that remain uncertain.

**Robust findings:** (i) The signs of all demographic distance coefficients are stable across pooled GLS, pair FE, PPML, and extensive margin logit. (ii) The KAOPEN interaction terms survive all clustering approaches (pair, reporter, two-way; all p < 0.04) and pair FE (all p < 0.001). (iii) The portfolio-but-not-FDI contrast holds in every specification. (iv) The geography of projected bilateral reallocations is invariant to GE adjustment.

**Uncertain findings:** (i) The precise rate-associated share (58%) is a statistical decomposition based on a generated regressor approach that may overstate precision; it should be read as indicative of a substantially larger bilateral rate share than the multilateral 9%, not as a precise point estimate. (ii) Level $\Delta Z$ coefficients become insignificant under conservative clustering and pair FE without interactions, reflecting the difficulty of separating slowly evolving demographics from other persistent pair characteristics. (iii) The GE clearing overlay is stylized and applies a uniform rate adjustment, which understates heterogeneity in country-specific rate responses.

Additional data limitations include: the CPIS data cover positions (stocks) rather than flows; the direct price control test is constrained to 506 OECD pairs with bilateral bond yield differentials; we do not model third-country effects; and the bilateral projections beyond 2030 use frozen GDP and constant KAOPEN.


# 10. Conclusion

This paper provides bilateral evidence that demographic structure predicts international capital flows. Using augmented gravity models on comprehensive CPIS and CDIS data, we show that bilateral demographic distance---the difference in age-structure polynomial variables between origin and destination---is highly significant for portfolio investment but not for FDI. The effect is concentrated in debt securities, amplified by financial openness, and robust to CCA exclusion. A companion channel decomposition confirms that approximately 88 percent of the demographic current account effect is direct rather than price-associated, consistent with the bilateral gravity finding that price controls have limited impact on the demographic flow effect.

Together, these results establish that demographics are a first-order determinant of the bilateral pattern of international portfolio investment. A two-stage statistical decomposition reveals that 58 percent of the bilateral demographic R² improvement is associated with yield differentials, with 42 percent through non-rate channels---a pattern consistent with interest rates determining the geography of demographic flows (bilateral) while aggregate volume is driven by direct savings adjustment (multilateral). Projections through 2050---incorporating WEO GDP growth and a general equilibrium rate clearing overlay---identify a "second wave" of demographic capital exporters: Korea, China, Spain, Italy, and Thailand. Even after GE damping of approximately 24 percent, the Korea--India pair is projected to nearly triple in bilateral portfolio holdings, driven by demographics and India's rapid GDP growth. The GE adjustment dampens the *volume* of cross-border flows but not their *geography*: the net reallocation pressure between countries is determined entirely by demographics. For a world aging unevenly, this implies persistent and growing demographic capital flows from old to young economies, statistically associated with both yield differentials and direct savings allocation, and moderated primarily by financial openness.


# References

Anderson, J. E., & van Wincoop, E. (2003). Gravity with gravitas: A solution to the border puzzle. *American Economic Review*, 93(1), 170--192.

Backus, D., Cooley, T., & Henriksen, E. (2014). Demography and low-frequency capital flows. *Journal of International Economics*, 92(S1), S94--S102.

Baron, R. M., & Kenny, D. A. (1986). The moderator--mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. *Journal of Personality and Social Psychology*, 51(6), 1173--1182.

Bertrand, M., Duflo, E., & Mullainathan, S. (2004). How much should we trust differences-in-differences estimates? *Quarterly Journal of Economics*, 119(1), 249--275.

Cameron, A. C., Gelbach, J. B., & Miller, D. L. (2011). Robust inference with multiway clustering. *Journal of Business & Economic Statistics*, 29(2), 238--249.

Carvalho, C., Ferrero, A., & Nechio, F. (2016). Demographics and real interest rates: Inspecting the mechanism. *European Economic Journal*, 60(1), 98--115.

Chinn, M. D., & Ito, H. (2006). What matters for financial development? Capital controls, institutions, and interactions. *Journal of Development Economics*, 81(1), 163--192.

Daude, C., & Fratzscher, M. (2007). The pecking order of cross-border investment. *Journal of International Economics*, 74(1), 94--119.

Fair, R. C., & Dominguez, K. M. (1991). Effects of the changing U.S. age distribution on macroeconomic equations. *American Economic Review*, 81(5), 1276--1294.

Gruber, J. W., & Kamin, S. B. (2007). Explaining the global pattern of current account imbalances. *Journal of International Money and Finance*, 26(4), 500--522.

Higgins, M. (1998). Demography, national savings, and international capital flows. *International Economic Review*, 39(2), 343--369.

International Monetary Fund. (2013). The External Balance Assessment (EBA) methodology (IMF Working Paper WP/13/272).

Koomen, M., & Wicht, L. (2020). Demographics and current account imbalances. Working Paper.

Krueger, D., & Ludwig, A. (2007). On the consequences of demographic change for rates of returns to capital, and the distribution of wealth and welfare. *Journal of Monetary Economics*, 54(1), 49--87.

Lucas, R. E. (1990). Why doesn't capital flow from rich to poor countries? *American Economic Review*, 80(2), 92--96.

Lane, P. R., & Milesi-Ferretti, G. M. (2007). The external wealth of nations mark II: Revised and extended estimates of foreign assets and liabilities, 1970--2004. *Journal of International Economics*, 73(2), 223--250.

Lane, P. R., & Milesi-Ferretti, G. M. (2008). International investment patterns. *Review of Economics and Statistics*, 90(3), 538--549.

Mayer, T., & Zignago, S. (2011). Notes on CEPII's distances measures: The GeoDist database (CEPII Working Paper 2011-25).

Obstfeld, M., & Rogoff, K. (2005). Global current account imbalances and exchange rate adjustments. *Brookings Papers on Economic Activity*, 2005(1), 67--123.

Portes, R., & Rey, H. (2005). The determinants of cross-border equity flows. *Journal of International Economics*, 65(2), 269--296.

Santos Silva, J. M. C., & Tenreyro, S. (2006). The log of gravity. *Review of Economics and Statistics*, 88(4), 641--658.

Tinbergen, J. (1962). *Shaping the World Economy: Suggestions for an International Economic Policy*. New York: Twentieth Century Fund.
